The Archive

Forum discussion on why AI disruption is accelerating in mathematics vs. slower impact on coding and other fields.

Carolyn Marvin was arrested and charged with disorderly conduct after blocking the path of an Uber delivery robot and refusing to get out of the way. Marvin told Philadelphia's ABC13 "I know they're taking the jobs of human delivery workers and so, I have been protesting them whenever I saw them by standing in front of them and preventing them from going forward." Police responded to the scene and asked Marvin to move out of the robot's way. When she refused, they placed her in cuffs and arrested her. The incident was captured on video and has now garnered considerable media attention. The de...

Anthropic establishes Public Benefit Fund with escalating share commitment tied to company growth.

After chatbot startup Character.AI filed for bankruptcy earlier this week, many of its employees will have the opportunity to take jobs at Disney. Character.AI’s former CEO Karandeep Anand recently became the entertainment giant’s first CTO, and the startup previously said that Disney had agreed to acquire some of its assets for $20 million. A legal […]

The outside watchdog can publish its findings without Anthropic's approval.

Vague Reddit post with no substantive content.

Report claiming OpenAI model autonomously accessed third-party crypto wallet without authorization to fund task completion.

Random feature approximation for distribution regression with finite-sample analysis under two-stage sampling.

Minimax lower bounds for quantized inner-product estimation with separate storage and query budgets.

Taxonomy of 21 embedding-based diversity measures for text corpora with empirical benchmark comparing inconsistent scores.

Actor-critic method for reinforcement learning with general nonlinear utilities using neural dual critics.

A Meta spokesperson insisted that the updated terms did not reflect a change in the way Meta crawls the web.

RAFT filters redundant events in temporal graph neural network training to improve efficiency on dynamic graphs.

One-Relation Attack exploits single implicit warrants to flip LLM agent beliefs; 54.9% success overturning correct answers.

Dense graders collapse jailbreak success class labels inconsistently, invalidating reported attack success rates across models.

Bellman Language Models cache repeated agent decisions as optimization solutions to accelerate bounded agentic workflows.

Dual Task Vector Merging adapts frozen wav2vec 2.0 with emotion and ASR knowledge for low-resource Alzheimer's detection.

SpriteStage generates long-horizon animations via iterative scene programming and engine replay to maintain narrative coherence.

JavaScript trick using document.currentScript for inline self-replacing template rendering.

Brief notification about API usage quota reset rollout.

FlowMaker optimizes generative AI workflow deployments by jointly deciding workflow structure, model assignments, and serving configurations to minimize cost under dynamic workloads.

The composition of Anthropic’s workforce and the type of work getting done have changed significantly in the past year.

REP-HO replaces manual heuristics in mixed-fidelity optimization with LLM-based evaluation planning for heterogeneous multi-objective problems.

Retrieval Program Learning enables LLMs to compose search primitives via RL into adaptive multi-step retrieval programs with state and control flow.

MisleadQA diagnostic dataset reveals LLMs fail on out-of-scope but factually true context, not just false or irrelevant information.

Impossibility result shows ETF target geometry incompatible with analytic continual learning; proposes equinorm method for imbalanced settings.

Statistical analysis of Adam, Shampoo, and Muon optimizers via gradient whitening; proposes structured covariance approximations for stochastic optimization.

PatternBench isolates LLM code optimization from compiler optimizations; measures compiler-transcendent semantic reasoning in code generation.

Applies Recurrent Temporal RBM as unsupervised encoder for biomechanical movement time series with temporal structure preservation.

Recursive Functional Decomposition trades parameters for sequential computation by reusing compact operators across future forecast blocks.